Hadoop configuration & performance tuning
139 likes41,392 views
This document provides an agenda and overview for a presentation on Hadoop 2.x configuration and MapReduce performance tuning. The presentation covers hardware selection and capacity planning for Hadoop clusters, key configuration parameters for operating systems, HDFS, and YARN, and performance tuning techniques for MapReduce applications. It also demonstrates the Hadoop Vaidya performance diagnostic tool.
Hadoop configuration & performance tuning
- 1. 1Pivotal Confidential–Internal Use Only 1 Hadoop 2.x Configuration & Map/Reduce Performance Tuning Suhas Gogate, Architect Hadoop Engg CF-Meetup, SFO (20th May 2014 )
- 2. A NEW PLATFORM FOR A NEW ERA
- 3. 3Pivotal Confidential–Internal Use Only About Me (https://www.linkedin.com/in/vgogate) Since 2008, active in Hadoop infrastructure and ecosystem components with lead Hadoop technology based companies – Yahoo, Netflix, Hortonworks, EMC-Greenplum/Pivotal Founder and PMC member/committer of the Apache Ambari project Contributed Apache “Hadoop Vaidya” – Performance diagnostics for M/R Prior to Hadoop, – IBM Almaden Research (2000-2008), CS Software & Storage systems. – In early days (1993) of my career, worked with a team that built first Indian super computer, PARAM (Transputer based MPP system) at Center for Development of Advance Computing (CDAC, Pune)
- 4. 4Pivotal Confidential–Internal Use Only Agenda Introduction to Hadoop 2.0 – HDFS, YARN, Map/Reduce (M/R) Hadoop Cluster – Hardware selection & Capacity Planning Key Hadoop Configuration Parameters – Operating System, Local FS, HDFS, YARN Performance Tuning of M/R Applications – Hadoop Vaidya demo (MAPREDUCE-3202)
- 5. 5Pivotal Confidential–Internal Use Only Introduction to Hadoop
- 6. 6Pivotal Confidential–Internal Use Only Hadoop 1.0 -> 2.0 Image Courtesy Arun Murthy, Hortonworks
- 7. 7Pivotal Confidential–Internal Use Only HDFS Architecture
- 8. 8Pivotal Confidential–Internal Use Only Hadoop Map/Reduce 1.0
- 9. 9Pivotal Confidential–Internal Use Only Hadoop YARN 2.0 Architecture
- 10. 10Pivotal Confidential–Internal Use Only Hardware selection & Capacity planning
- 11. 11Pivotal Confidential–Internal Use Only Hadoop Cluster: Hardware selection Hadoop runs on cluster of commodity machines – Does NOT mean unreliable & low-cost hardware – 2008: ▪ Single/Dual socket, 1+ GHz, 4-8GB RAM, 2-4 Cores, 2-4 1TB SATA drives, 1GbE NIC – 2014+: ▪ Dual socket 2+ GHz, 48-64GB RAM, 12-16 Cores, 12-16 1/3TB SATA drives, 2-4 bonded NICs Hadoop cluster comprises separate h/w profiles for, – Master nodes ▪ More reliable & available configuration ▪ Low storage/memory requirements compared to worker/data nodes (except HDFS Name node) – Worker nodes ▪ Not necessarily less reliable although not configured for high availability (e.g. JBOD not RAID) ▪ Resource requirements on storage, memory, CPU & network b/w are based on workload profile See Appliance/Reference architectures from EMC/IBM/Cisco/HP/Dell etc
- 12. 12Pivotal Confidential–Internal Use Only Cluster capacity planning Initial cluster capacity is commonly based on HDFS data size & growth projection – Keep data compression in mind Cluster size – HDFS space per node = (Raw disk space per node – 20/25% non-DFS local storage) / 3 (RF) – Cluster nodes = Total HDFS space / HDFS space per node Start with balanced individual node configuration in terms of CPU/Memory/Number of disks – Keep provision for growth as you learn more about the workload – Guidelines for individual worker node configuration ▪ Latest generation processor(s) with 12/16 cores total, is a reasonable start ▪ 4 to 6 GB memory per core ▪ 1/1.5 disks per core ▪ 1 to 3 TB SATA disks per core ▪ 1GbE NIC Derive resource requirements service roles (typically Name node) – NN -> 4-6 cores, default 2GB + 1 GB roughly every 100TB of raw disk space, 1M Objects – RM/NN/DN -> 2GB RAM, 1 core (thumb rule)
- 13. 13Pivotal Confidential–Internal Use Only Ongoing Capacity Planning Workload profiling – Make sure applications running on the cluster are well tuned to utilize cluster resources – Gather average CPU/IO/Mem/Disk utilization stats across worker nodes – Identify the resource bottlenecks for your workload and provision accordingly ▪ E,g, bump up more cores per node if CPU is bottleneck compared to other resources such as storage, I/O bandwidth ▪ E.g. If memory is a bottleneck (i.e. low utilization on CPU and I/O), then add more memory per node to let more tasks run per node. – HDFS storage growth rate & current capacity utilization Latency sensitive applications – Do planned project on-boarding ▪ Estimate resource requirement ahead of time (Capacity Calculator) – Use Resource Scheduler ▪ Schedule the jobs to maximize the resource utilization over the time
- 14. 14Pivotal Confidential–Internal Use Only Hadoop Cluster Configuration
- 15. 15Pivotal Confidential–Internal Use Only Key Hadoop Cluster Configuration – OS Operating System (RHEL/CentOS 6.1+) – Mount disk volumes with NOATIME (speed up reads) – Disable transparent Huge page compaction ▪ # echo never > /sys/kernel/mm/redhat_transparent_hugepages/defrag – Turn off caching on disk controller – vm.swappiness = 0 – vm.overcommit_memory = 1 – vm.overcommit_ratio = 100 – net.core.somaxconn=1024 (default socket listen queue size 128) – Choice of Linux I/O scheduler Local File System – Ext3 (reliable and recommended) vs Ext4 vs XFS – Default max open file descriptors per user (default 1K may change to 32K) – Reduce FS reserve blocks space (default 5% -> 0% on non-os partitions) BIOS – Disable BIOS power saving options may boost node performance
- 16. 16Pivotal Confidential–Internal Use Only Key Hadoop Cluster Configuration - HDFS Use multiple disk mount points – dfs.datanode.data.dir (use all attached disks to data node) – dfs.namenode.name.dir (NN metadata redundancy on disk) DFS Block size – 128MB (you can override it while writing new files with different block size) Local file system buffer – io.file.buffer.size = 131072 (128KB) – Io.sort.factor = 50 to 100 (number merge streams while sorting file) NN/DN concurrency – dfs.namenode.handler.count (100) – dfs.datanode.max.transfer.threads (4096) Datanode Failed volumes tolerated – dfs.datanode.failed.volumes.tolerated
- 17. 17Pivotal Confidential–Internal Use Only Key Hadoop Cluster Configuration - HDFS Short circuit read – dfs.client.read.shortcircuit = true – dfs.domain.socket.path JVM options (hadoop-env.sh) – export HADOOP_NAMENODE_OPTS=” ▪ -Dcom.sun.management.jmxremote -Xms${dfs.namenode.heapsize.mb}m ▪ -Xmx${dfs.namenode.heapsize.mb}m ▪ -Dhadoop.security.logger=INFO,DRFAS ▪ -Dhdfs.audit.logger=INFO,RFAAUDIT ▪ -XX:ParallelGCThreads=8 ▪ -XX:+UseParNewGC -XX:+UseConcMarkSweepGC ▪ -XX:+HeapDumpOnOutOfMemoryError -XX:ErrorFile=${HADOOP_LOG_DIR}/hs_err_pid%p.log $HADOOP_NAMENODE_OPTS” – export HADOOP_DATANODE_OPTS=” ▪ -Dcom.sun.management.jmxremote ▪ -Xms${dfs.datanode.heapsize.mb}m ▪ -Xmx${dfs.datanode.heapsize.mb}m ▪ -Dhadoop.security.logger=ERROR,DRFAS $HADOOP_DATANODE_OPTS"
- 18. 18Pivotal Confidential–Internal Use Only Key Hadoop Cluster Configuration - YARN Use multiple disk mount points on the worker node – yarn.nodemanager.local-dirs – yarn.nodemanager.log-dirs Memory allocated for node manager containers – yarn.nodemanager.resource.memory-mb ▪ Total memory on worker node allocated for all the containers running in parallel – yarn.scheduler.minimum-allocation-mb ▪ Minimum memory requested for map/reduce task container. ▪ Based on CPU cores and available memory on the node, this parameter can limit number of max containers per node – yarn.nodemanager.maximum-allocation-mb ▪ Default to yarn.nodemanager.resource.memory-mb – yarn.mapreduce.map.memory.mb, yarn.mapreduce.reduce.memory.mb ▪ Default values for map/reduce task container memory. User can override them through job configuration – yarn.nodemanager.vmem-pmem-ratio = 2.1
- 19. 19Pivotal Confidential–Internal Use Only Key Hadoop Cluster Configuration - YARN Use Yarn log aggregation – yarn.log-aggregation-enable RM/NM JVM options (yarn-env.sh) – export YARN_RESOURCEMANAGER_HEAPSIZE=2GB – export YARN_NODEMANAGER_HEAPSIZE=2GB – YARN_OPTS="$YARN_OPTS -server – -Djava.net.preferIPv4Stack=true – -XX:+UseParNewGC -XX:+UseConcMarkSweepGC – -XX:+HeapDumpOnOutOfMemoryError – -XX:ErrorFile=${YARN_LOG_DIR}/hs_err_pid%p.log"
- 20. 20Pivotal Confidential–Internal Use Only Hadoop Configuration Advisor - Tool Given – Data size, Growth rate, Workload profile, Latency/QoS requirements Suggests – Capacity requirement for Hadoop cluster (reasonable starting point) ▪ Resource requirements for various services roles ▪ Hardware profiles for master/worker nodes (No specific h/w vendor ) – Cluster services topology i.e. placement of service roles to nodes – Optimal services configuration for given hardware specs
- 21. 21Pivotal Confidential–Internal Use Only Performance Tuning of M/R applications
- 22. 22Pivotal Confidential–Internal Use Only Hadoop Map/Reduce - WordCount
- 23. 23Pivotal Confidential–Internal Use Only Optimizing M/R applications – key features Speculative Execution Use of Combiner Data Compression – Intermediate: LZO(native)/Snappy, Output: BZip2, Gzip Avoid map side disk spills – io.sort.mb Increased replication factor for out-of-band Hdfs access Distributed cache Map output partitioner Appropriate granularity for M/R tasks – Mapreduce.map.minsplitsize – Mapreduce.map.maxsplitsize – Optimal number of reducers
- 24. 24Pivotal Confidential–Internal Use Only Performance Benchmark – Teragen Running teragen out-of-box will not utilize the cluster hardware resources Determine number of map tasks to run on each node to exploit max I/O bandwidth – Depends on number of disks on each node Example – 10 nodes, 5 disks/node, per node memory for m/r tasks 50GB – Hadoop jar hadoop-mapreduce-examples-2.x.x.jar teragen – -Dmapred.map.tasks=50 -Dmapreduce.map.memory.mb=10GB – 10000000000 /teragenoutput
- 25. 25Pivotal Confidential–Internal Use Only Hadoop M/R Benchmark – Terasort Terasort Example hadoop jar hadoop-mapreduce/hadoop-mapreduce-examples-2.x.x.jar terasort – -Ddfs.replication=1 -Dmapreduce.task.io.sort.mb=500 – -Dmapreduce.map.sort.spill.percent=0.9 – -Dmapreduce.reduce.shuffle.parallelcopies=10 – -Dmapreduce.reduce.shuffle.memory.limit.percent=0.1 – -Dmapreduce.reduce.shuffle.input.buffer.percent=0.95 – -Dmapreduce.reduce.input.buffer.percent=0.95 – -Dmapreduce.reduce.shuffle.merge.percent=0.95 – -Dmapreduce.reduce.merge.inmem.threshold=0 – -Dmapreduce.job.speculative.speculativecap=0.05 – -Dmapreduce.map.speculative=false – -Dmapreduce.reduce.speculative=false -Dmapreduce.job.jvm.numtasks=-1
- 26. 26Pivotal Confidential–Internal Use Only Hadoop M/R Benchmark – Terasort – -Dmapreduce.job.reduces=84 -Dmapreduce.task.io.sort.factor=100 – -Dmapreduce.map.output.compress=true – -Dmapreduce.map.output.compress.codec= ▪ org.apache.hadoop.io.compress.SnappyCodec – -Dmapreduce.job.reduce.slowstart.completedmaps=0.4 – -Dmapreduce.reduce.merge.memtomem.enabled=false – -Dmapreduce.reduce.memory.totalbytes=12348030976 – -Dmapreduce.reduce.memory.mb=12288 – -Dmapreduce.reduce.java.opts= ▪ "-Xms11776m -Xmx11776m -XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode - XX:+CMSIncrementalPacing -XX:ParallelGCThreads=4" – -Dmapreduce.map.memory.mb=4096 – -Dmapreduce.map.java.opts="-Xmx1356m" – /terasort-input /terasort-output
- 27. 27Pivotal Confidential–Internal Use Only Hadoop Vaidya: Performance diagnostic tool
- 28. 28Pivotal Confidential–Internal Use Only Hadoop Vaidya: Rule based performance diagnostic tool • Rule based performance diagnosis of M/R jobs – Set of pre-defined diagnostic rules – Diagnostic rules execution against job config & Job History logs – Targeted advice for discovered problems • Extensible framework – You can add your own rules, • Based on a rule template and published job counters – Write complex rules using existing simpler rules Vaidya: An expert (versed in his own profession , esp. in medical science) , skilled in the art of healing , a physician
- 29. 29Pivotal Confidential–Internal Use Only Hadoop Vaidya: Diagnostic Test Rule <DiagnosticTest> <Title>Balanaced Reduce Partitioning</Title> <ClassName> org.apache.hadoop.vaidya.postexdiagnosis.tests.BalancedReducePartitioning </ClassName> <Description> This rule tests as to how well the input to reduce tasks is balanced </Description> <Importance>High</Importance> <SuccessThreshold>0.40</SuccessThreshold> <Prescription>advice</Prescription> <InputElement> <PercentReduceRecords>0.85</PercentReduceRecords> </InputElement> </DiagnosticTest>
- 30. 30Pivotal Confidential–Internal Use Only Hadoop Vaidya: Report Element <TestReportElement> <TestTitle>Balanaced Reduce Partitioning</TestTitle> <TestDescription> This rule tests as to how well the input to reduce tasks is balanced </TestDescription> <TestImportance>HIGH</TestImportance> <TestResult>POSITIVE(FAILED)</TestResult> <TestSeverity>0.69</TestSeverity> <ReferenceDetails> * TotalReduceTasks: 4096 * BusyReduceTasks processing 0.85% of total records: 3373 * Impact: 0.70 </ReferenceDetails> <TestPrescription> * Use the appropriate partitioning function * For streaming job consider following partitioner and hadoop config parameters * org.apache.hadoop.mapred.lib.KeyFieldBasedPartitioner * -jobconf stream.map.output.field.separator, -jobconf stream.num.map.output.key.fields </TestPrescription> </TestReportElement>
- 31. 31Pivotal Confidential–Internal Use Only Hadoop Vaidya: Example Rules Balanced Reduce Partitioning – Check if intermediate data is well partitioned among reducers. Map/Reduce tasks reading HDFS files as side effect – Checks if HDFS files are being read as side effect and causing the access bottleneck across map/reduce tasks Percent Re-execution of Map/Reduce tasks Granularity of Map/Reduce task execution Map tasks data locality – This rule detects % data locality for Map tasks Use of Combiner & Combiner efficiency – Checks if there is a potential in using combiner after map stage Intermediate data compression – Checks if intermediate data is compressed to lower the shuffle time
- 32. 32Pivotal Confidential–Internal Use Only Hadoop Vaidya: Demo Thank you! • For Hadoop Vaidya demo • Download Pivotal HD single node VM (https://network.gopivotal.com/products/pivotal-hd) • Run M/R job • After job completed, go Resource Manager UI • Select Job History • To look at the diagnostic report click on “Vaidya Report” link in left menu • Vaidya Patch submitted to Apache Hadoop • https://issues.apache.org/jira/browse/MAPREDUCE-3202
- 33. A NEW PLATFORM FOR A NEW ERA